A novel method of using neural networks to predict wine composition

Area of Science:

• Analytical Chemistry
• Machine Learning
• Spectroscopy

Background:

• Wine quality relies on precise chemical composition analysis.
• Current methods for wine component analysis can be expensive and time-consuming.
• Infrared (IR) spectroscopy offers a rapid, non-destructive analytical technique.

Purpose of the Study:

• To develop and validate a machine learning-based method for diagnosing key wine components using IR absorption spectra.
• To compare the performance of Partial Least Squares (PLS) regression, perceptrons, and convolutional neural networks (CNNs) for wine analysis.
• To assess the feasibility of this method as a cost-effective alternative to existing commercial solutions for rapid wine analysis.

Main Methods:

• Training machine learning models (PLS, perceptrons, CNNs) on IR spectra of simulated white and red wines (1734 samples).
• Applying the trained models to analyze IR spectra of real wines.
• Quantifying concentrations of ethanol, sugars, acids, glycerol, and sulfur dioxide.

Main Results:

• The developed method achieved accuracy comparable to commercial methods for ethanol and sugars, and slightly inferior for other components.
• Convolutional neural networks (CNNs) demonstrated superior performance for determining ethanol and sugar concentrations compared to PLS and perceptrons.
• All tested models provided similar accuracy for acids, glycerol, and sulfur dioxide determination.

Conclusions:

• The proposed IR spectroscopy and machine learning approach provides accurate wine component analysis.
• CNNs show particular promise for quantifying ethanol and sugars in wine.
• This method offers a significant reduction in software development costs for rapid wine analysis instruments, presenting a viable alternative to commercial systems.